Multimode process monitoring using prototype-based Gaussian mixture model

Modern industrial processes often have multiple operating modes because of their complexity and manufacturing strategy changes. Meanwhile, the within-mode process data can also be nonlinear and non-Gaussian distributed. To deal with the problems above, Gaussian mixture model (GMM) has recently been applied to multimode process monitoring which achieves better performance than traditional multivariate statistical process monitoring techniques. However, the training of GMM usually depends on expectation-maximization (EM) algorithm, which is sensitive to initial values and requires a priori the number of components. To alleviate these problems above, this paper proposes a new algorithm to train GMM based on prototype selection, named prototype-based GMM (PGMM), and applies it to multimode process monitoring. The algorithm can determine the number of Gaussian components adaptively and is highly efficient and stable. The parameters in the algorithm need not to be specially tuned because of their clear statistical meanings. Through experiments of a numeric example and the TE benchmark problem, the effectiveness of the proposed method is demonstrated.